Research map / Video / Action recognition
Large-Scale Weakly-Supervised Pre-Training for Video Action Recognition
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pre-train on 65M hashtag-labeled web videos | R(2+1)D-34 on Kinetics · Kinetics | top-1 accuracy: 67.0% -> 74.8% | Helped |
| Increase weakly supervised pre-training data from Kinetics to 65M videos | R(2+1)D-34 · EPIC-Kitchens | accuracy: 12.3% -> 16.1% | Helped |
| Increase pre-training data volume (ablation) | R(2+1)D-34 · Kinetics | accuracy | Helped |
| Use tail-preserving rather than random data sampling (ablation) | R(2+1)D-34 fc-only · Kinetics | accuracy | Helped |
| Increase model capacity from R(2+1)D-18 to R(2+1)D-152 (ablation) | R(2+1)D pre-trained on IG-Kinetics-65M · Kinetics | top-1 accuracy: 76.0 -> 79.9 | Helped |
| Increase capacity from R(2+1)D-101 to R(2+1)D-152 (ablation) | R(2+1)D pre-trained on IG-Kinetics-65M · EPIC-Kitchens | top-1 accuracy: 24.9 -> 23.7 | Hurt |
| Use pre-training labels with greater overlap with the target labels (ablation) | R(2+1)D-34 · Kinetics | accuracy: at least 5.5% | Helped |
| Use Verb+Noun labels instead of a less skewed target-overlapping label set (ablation) | R(2+1)D-34 · EPIC-Kitchens actions | accuracy | Hurt |
| Increase Verb+Noun pre-training label count up to 1350 (ablation) | R(2+1)D-34 · Kinetics | accuracy: ~1% | Helped |
| Increase Verb+Noun pre-training label count beyond 1350 (ablation) | R(2+1)D-34 full-ft · Kinetics | accuracy | No clear effect |
| Reduce Kinetics pre-training label count (ablation) | R(2+1)D-34 · Kinetics | accuracy | Hurt |
| Use long rather than short videos with the same 5M-video budget (ablation) | R(2+1)D-34 · Kinetics | accuracy: 57.4 -> 60.6; 3.2% | Helped |
Closest papers by meaning
- Learning from Weakly-Labeled Web Videos via Exploring Sub-concepts (AAAI 2022 · similarity 0.93)
- VideoNet: A Large-Scale Dataset for Domain-Specific Action Recognition (CVPR 2026 · similarity 0.90)
- Semi-Weakly-Supervised Learning of Complex Actions From Instructional Task Videos (CVPR 2022 · similarity 0.90)
- Action Detail Matters: Refining Video Recognition with Local Action Queries (CVPR 2025 · similarity 0.88)
- Semi-supervised Active Learning for Video Action Detection (AAAI 2024 · similarity 0.88)
- Stable Mean Teacher for Semi-supervised Video Action Detection (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.